Hi Serge, I made some experiments with GPU computing with JavaScript. I understood that the computings are made forever inside the GPU, so you just throw data and programs once and you let the GPU compute for you. Yet the problem is to read the data once they are computed. With webGL it seems impossible, with webCL it is possible and not easy. Hence the choice of openCL I guess. I relate here how I could compute the powers of a Markov matrix here: http://revue.sesamath.net/spip.php?article651 (especially click on "webGL sans three.js"). I also made some experiments here: http://irem.univ-reunion.fr/spip.php?article797 (but they use three.js if I remember well) Happy readings, and, yes, I feel interersted in these subjects Alain On Tue, May 12, 2015 at 7:48 PM, Serge Stinckwich < serge.stinckwich@gmail.com> wrote:
Dear all,
just to let you know, Cheikhou (in CC) is starting a student internship in my lab. He will work on Epidemiology Modelling with KENDRICK: http://ummisco.github.io/kendrick/ the platform that we are developing in order to analyse and visualise diseases models behaviours.
We would like first to implement a GPU version of the Gillespie Stochastic Simulation Algorithm (GSSA): http://en.wikipedia.org/wiki/Gillespie_algorithm and after that also implement SPH simulations: https://en.wikipedia.org/wiki/Smoothed-particle_hydrodynamics
We are looking at the code of OpenCL and VirtualGPU done by Ronnie. What we have understand until now : - OpenCL package : low-level stuff to be able to interface OpenCL kernels with Pharo - VirtualGPU: high-level API on top of OpenCL in order to ease the task of people who wants to use OpenCL. VirtualGPU provide high-level operations on matrix and image at the moment.
@Ronie: What is not clear at the moment in our mind : when you build a VirtualGPU program with the DSL, do you have the overhead of communications every time you execute a VirtualGPU instruction or all the the instructions are sent at the same time and run on the GPU ?
In our context, for building a GSSA algorithm, I guess we just have to combine same VGPU instructions (matrix computations) but for doing SPH simulations, we will have to provide our own instructions. Is there any documentation in order to add own kernel and instructions ?
I know that others guys at INRIA (Stéphane ?) are interested by GPU. Is it possible to join our effort to share what we are doing ?
Regards, -- Serge Stinckwich UCBN & UMI UMMISCO 209 (IRD/UPMC) Every DSL ends up being Smalltalk http://www.doesnotunderstand.org/